ClickHouse和他的朋友们（13）ReplicatedMergeTree表引擎及同步机制

原文出处：https://bohutang.me/2020/09/13/clickhouse-and-friends-replicated-merge-tree/

最后更新: 2020-09-13

在 MySQL 里，为了保证高可用以及数据安全性会采取主从模式，数据通过 binlog 来进行同步。

在 ClickHouse 里，我们可以使用 ReplicatedMergeTree 引擎，数据同步通过 zookeeper 完成。

本文先从搭建一个多 replica 集群开始，然后一窥底层的机制，简单吃两口。

1. 集群搭建

搭建一个 2 replica 测试集群，由于条件有限，这里在同一台物理机上起 clickhouse-server(2个 replica) + zookeeper(1个)，为了避免端口冲突，两个 replica 端口会有所不同。

1.1 zookeeper

docker run  -p 2181:2181 --name some-zookeeper --restart always -d zookeeper

1.2 replica集群

replica-1 config.xml:

   <zookeeper>
      <node index="1">
         <host>172.17.0.2</host>
         <port>2181</port>
      </node>
   </zookeeper>

   <remote_servers>
      <mycluster_1>
         <shard_1>
            <internal_replication>true</internal_replication>
            <replica>
               <host>s1</host>
               <port>9000</port>
            </replica>
            <replica>
               <host>s2</host>
               <port>9001</port>
            </replica>
         </shard_1>
      </mycluster_1>
   </remote_servers>

   <macros>
      <cluster>mycluster_1</cluster>
      <shard>1</shard>
      <replica>s1</replica>
   </macros>


   <tcp_port>9101</tcp_port>
   <interserver_http_port>9009</interserver_http_port>
   <path>/cluster/d1/datas/</path>

replica-2 config.xml:

   <zookeeper>
      <node index="1">
         <host>172.17.0.2</host>
         <port>2181</port>
      </node>
   </zookeeper>

   <remote_servers>
      <mycluster_1>
         <shard_1>
            <internal_replication>true</internal_replication>
            <replica>
               <host>s1</host>
               <port>9000</port>
            </replica>
            <replica>
               <host>s2</host>
               <port>9001</port>
            </replica>
         </shard_1>
      </mycluster_1>
   </remote_servers>

   <macros>
      <cluster>mycluster_1</cluster>
      <shard>1</shard>
      <replica>s2</replica>
   </macros>

   <tcp_port>9102</tcp_port>
   <interserver_http_port>9010</interserver_http_port>
   <path>/cluster/d2/datas/</path>

1.3 创建测试表

CREATE TABLE default.rtest1 ON CLUSTER 'mycluster_1'
(
    `id` Int64,
    `p` Int16
)
ENGINE = ReplicatedMergeTree('/clickhouse/tables/replicated/test', '{replica}')
PARTITION BY p
ORDER BY id

1.4 查看 zookeeper

docker exec -it some-zookeeper bash
./bin/zkCli.sh

[zk: localhost:2181(CONNECTED) 17] ls /clickhouse/tables/replicated/test/replicas
[s1, s2]

两个 replica 都已经注册到 zookeeper。

2. 同步原理 

如果在 replica-1 上执行了一条写入:

replica-1> INSERT INTO rtest VALUES(33,33);

数据是如何同步到 replica-2 的呢？

s1.  replica-1> StorageReplicatedMergeTree::write --> ReplicatedMergeTreeBlockOutputStream::write(const Block & block)
s2.  replica-1> storage.writer.writeTempPart，写入本地磁盘
s3.  replica-1> ReplicatedMergeTreeBlockOutputStream::commitPart
s4.  replica-1> StorageReplicatedMergeTree::getCommitPartOp，提交LogEntry到zookeeper，信息包括:
    ReplicatedMergeTreeLogEntry {
     type: GET_PART,
     source_replica: replica-1,
     new_part_name: part->name,
     new_part_type: part->getType
    }
s5.  replica-1> zkutil::makeCreateRequest(zookeeper_path + "/log/log-0000000022")，更新log_pointer到zookeeper

s6.  replica-2> StorageReplicatedMergeTree::queueUpdatingTask()，定时pull任务
s7.  replica-2> ReplicatedMergeTreeQueue::pullLogsToQueue ，拉取
s8.  replica-2> zookeeper->get(replica_path + "/log_pointer") ，向zookeeper获取当前replica已经同步的位点
s9.  replica-2> zookeeper->getChildrenWatch(zookeeper_path + "/log") ，向zookeeper获取所有的LogEntry信息
s10. replica-2> 根据同步位点log_pointer从所有LogEntry中筛选需要同步的LogEntry，写到queue
s11. replica-2> StorageReplicatedMergeTree::queueTask，消费queue任务
s12. replica-2> StorageReplicatedMergeTree::executeLogEntry(LogEntry & entry)，根据LogEntry type执行消费
s13. replica-2> StorageReplicatedMergeTree::executeFetch(LogEntry & entry) 
s14. replica-2> StorageReplicatedMergeTree::fetchPart，从replica-1的interserver_http_port下载part目录数据
s15. replica-2> MergeTreeData::renameTempPartAndReplace，把文件写入本地并更新内存meta信息
s16. replica-2> 数据同步完成

也可以进入 zookeeper docker 内部直接查看某个 LogEntry:

[zk: localhost:2181(CONNECTED) 85] get /clickhouse/tables/replicated/test/log/log-0000000022
format version: 4
create_time: 2020-09-13 16:39:05
source replica: s1
block_id: 33_2673203974107464807_7670041793554220344
get
33_2_2_0

3. 总结

本文以写入为例，从底层分析了 ClickHouse ReplicatedMergeTree 的工作原理，逻辑并不复杂。

不同 replica 的数据同步需要 zookeeper(目前社区有人在做etcd的集成 pr#10376(https://github.com/ClickHouse/ClickHouse/pull/10376))做元数据协调，是一个订阅/消费模型，涉及具体数据目录还需要去相应的 replica 通过 interserver_http_port 端口进行下载。

replica 的同步都是以文件目录为单位，这样就带来一个好处：我们可以轻松实现 ClickHouse 的存储计算分离，多个 clickhouse-server 可以同时挂载同一份数据进行计算，而且这些 server 每个节点都是可写，虎哥已经实现了一个可以 work 的原型，详情请参考下篇 <存储计算分离方案与实现>。

4. 参考

[1] StorageReplicatedMergeTree.cpp (https://github.com/ClickHouse/ClickHouse/blob/f37814b36754bf11b52bd9c77d0e15f4d1825033/src/Storages/StorageReplicatedMergeTree.cpp)

[2] ReplicatedMergeTreeBlockOutputStream.cpp (https://github.com/ClickHouse/ClickHouse/blob/f37814b36754bf11b52bd9c77d0e15f4d1825033/src/Storages/MergeTree/ReplicatedMergeTreeBlockOutputStream.cpp)

[3] ReplicatedMergeTreeLogEntry.cpp (https://github.com/ClickHouse/ClickHouse/blob/f37814b36754bf11b52bd9c77d0e15f4d1825033/src/Storages/MergeTree/ReplicatedMergeTreeLogEntry.cpp)

[4] ReplicatedMergeTreeQueue.cpp (https://github.com/ClickHouse/ClickHouse/blob/f37814b36754bf11b52bd9c77d0e15f4d1825033/src/Storages/MergeTree/ReplicatedMergeTreeQueue.cpp)

全文完。

Enjoy ClickHouse :)

叶老师的「MySQL核心优化」大课已升级到MySQL 8.0，扫码开启MySQL 8.0修行之旅吧